RecExplainer: Aligning Large Language Models for Explaining Recommendation Models

This is the Repo for RecExplainer: Aligning Large Language Models for Recommendation Model Interpretability, which leverages LLMs as surrogate models for explaining black-box recommender models.

Introduction

RecExplainer is a framework designed to align LLMs for understanding and explaining target recommender models. This framework comprises two main components：The target recommender model, which is used to recommend items according to user preferences; The LLM, which serves as a surrogate model to mimick the behavior of the target recommender model and then explain such behavior in natural language.

To utilize the RecExplainer framework effectively, one must begin by training a target recommender model. This model not only serves as the black-box model to be explained, but also plays a crucial role in generating training and testing datasets for the LLM. We have crafted a suite of tasks to make the LLM better aligned with the predictive patterns of the target recommender model as well as the domain-specific knowledge: next item retrieval (task1), item ranking (task2), interest classification (task3), item discrimination (task4), history reconstruction (task6).

For training the LLM, we introduce three distinct alignment approaches:

1. Behavior alignment (RecExplainer-B): Operate in the language space, representing user preferences and item information as text to mimic the target model’s behavior.
2. Intention alignment (RecExplainer-I): Work in the latent space of the recommendation model, using user and item representations to understand the model’s behavior.
3. Hybrid alignment (RecExplainer-H): Combine both language and latent spaces.

Our evaluation of the RecExplainer framework is two-fold:

1. Alignment effect: Reflect the extent to which LLM understands the neurons and predictive pattern in the target model.
2. Explanation Generation Ability
   • Overall ratings: We use both GPT4 and human experts to annotate the quality of generated explanations.
   • Distinction and Coherence: We also train a classifier and a score predictor to further verify whether RecExplainer are indeed explaining its own predictions.

Environment Setting

Environment

conda create -n recexplainer python=3.9
conda activate recexplainer
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
pip install -r requirements.txt

Set OpenAI API Environment

If you want to use OpenAI API, you need to firstly run the following scripts in your console. If it is not Azure OpenAI API (OPENAI_API_TYPE is not "azure"), you only need to specify OPENAI_API_KEY and ENGINE.

export OPENAI_API_KEY_01=xxx;
export OPENAI_API_BASE_01=https://xxx.openai.azure.com/;
export OPENAI_API_VERSION_01=2023-03-15-preview;
export OPENAI_API_TYPE_01=azure;
export ENGINE_01=xxx;


###If you want to use multiple keys at the same time to speed up data generation.
export OPENAI_API_KEY_02=xxx;
export OPENAI_API_BASE_02=https://xxx.openai.azure.com/;
export OPENAI_API_VERSION_02=2023-03-15-preview;
export OPENAI_API_TYPE_02=azure;
export ENGINE_02=xxx;

Dataset Preparation for Target Recommender Model

For data preparation, you need to download three raw files: Amazon review, Amazon metadata, ShareGPT

• Amazon Video Games 5-core reviews: https://jmcauley.ucsd.edu/data/amazon_v2/categoryFilesSmall/Video_Games_5.json.gz
• Amazon Video Games metadata: https://jmcauley.ucsd.edu/data/amazon_v2/metaFiles2/meta_Video_Games.json.gz
• ShareGPT: https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/blob/main/ShareGPT_V3_unfiltered_cleaned_split.json

Then, run the following script to prepare data for training your target recommender models.

bash shell/preprocess_recmodel.sh

Clone the UniRec repo to $HOME/UniRec and install another virtual environment for UniRec, then run the following script to convert the data format according to UniRec requirements.

bash shell/unirec_prepare_data.sh

Training and Inference with the Target Recommender Model

Training

bash shell/unirec_train.sh

Inference

You need to copy some files to the UniRec directory in advance.

cp preprocess/unirec_utils/data4Exp.py $HOME/UniRec/unirec/main
cp $HOME/RecExplainer/data/unirec_raw_data/amazon_video_games_v3/train_ids.csv $HOME/UniRec/data/amazon_video_games_v3
cp $HOME/RecExplainer/data/unirec_raw_data/amazon_video_games_v3/test_ids.csv $HOME/UniRec/data/amazon_video_games_v3

bash shell/unirec_infer.sh

Dataset Preparation for RecExplainer Model

bash shell/recexplainer_data_pipeline.sh

After running the above script, you will get the following training and testing files:

For alignmen tasks

• behaviour_train.json & behaviour_valid.json: data for training and testing RecExplainer-B (behavior alignmet), also used to test the alignment performance of open source LLMs.
• intention_train.json & intention_valid.json: data for training and testing RecExplainer-I (intention alignment).
• both_train.json & both_valid.json: data for training and testing RecExplainer-H (hybrid alignment).

For explanation tasks

• explan_behaviour_valid.json: prompts for RecExplainer-B to generate explanations.
• explan_intention_valid.json: prompts for RecExplainer-I to generate explanations.
• explan_both_valid.json: prompts for RecExplainer-H to generate explanations.
• explan_chatgpt.csv: prompts for ChatGPT to generate explanations.

Train RecExplainer

bash shell/train.sh

Important Parameters:

• --data_names: the json file name with suffix ('_train.json' or '_valid.json') removed.
• --task_type: alignment method of the training stage. "behaviour" means behavior aligment, "intention" means intention aligment, "both" means hybrid aligment.

Evaluation

Alignment Effect

bash shell/infer_alignment.sh

Parameters:

• --data_names: the json file name with suffix ('_train.json' or '_valid.json') removed.
• --task_type: alignment method used by the model. "behaviour" means behavior aligment, "intention" means intention aligment, "both" means hybrid aligment, "none" means using LLM without alignment training.
• --inference_mode: the name of the inference task. "uid2hist" means history reconstruction task, "uid2next" means next item retrieval task, "uidiid2rank" means item ranking task, "uidiid2binary" means interest classification task.

Explanation Generation Ability

Generating Explanations with Aligned RecExplainer Model

bash shell/infer_explan.sh

Parameters:

• --data_names: the json file name with suffix ('_train.json' or '_valid.json') removed.
• --task_type: alignment method used by the model. "behaviour" means behavior aligment, "intention" means intention aligment, "both" means hybrid aligment, "none" means using LLM without alignment training.
• --inference_mode: the name of the inference task. "case study" means generating explanation texts.

After using different models to generate explanation texts, we can evaluate the quality of the explanations and finally compare the performance of different models on explanation generation task.

Overall ratings

We use GPT4 to score the explanations generated by each model.

bash shell/eval_explan.sh

Parameters:

• --model_names: Models used to generate explanations. Use commas to separate different model names.
• --model_response_files: Explanation Files generated by each model. Use commas to separate different files.
• --judge_query_file: Input file of gpt4
• --judge_response_file: Output file of gpt4

Distinction and Coherence

We train a classifier and a score predictor to further verify whether RecExplainer are indeed explaining its own predictions.

Data generation

bash discriminator/data_gen.sh

Train a classifier

The classifier aims to verify whether RecExplainer’s explanations are distinct from other LLM’s explanations.

bash discriminator/run_cls.sh

Train a score predictor

The score predictor aims to verify whether RecExplainer’s explanations reflect RecExplainer’s predictions.

bash discriminator/run_reg.sh

Citation

If you find this project useful in your research, please consider citing:

@article{lei2023recexplainer,
  title={RecExplainer: Aligning Large Language Models for Recommendation Model Interpretability},
  author={Lei, Yuxuan and Lian, Jianxun and Yao, Jing and Huang, Xu and Lian, Defu and Xie, Xing},
  journal={arXiv preprint arXiv:2311.10947},
  year={2023}
}